Jev demonstrator

connecting System One models leaf Jev API docs

Most of what your agents do is deciding, not writing.

Jev answers typed questions about a piece of state with calibrated probabilities, in one round trip, for a few thousandths of a cent. It cannot produce text and it cannot answer outside the options you declared. Your code branches on the values directly.

choice one of up to 255 options, probability per option score fractional position on a 2 to 10 level rubric noul calibrated yes/no, 0 to 1
Decisions run
0
this session
Questions answered
0
all in parallel
Median latency
–
live calls only
Total cost
$0
input tokens at $0.42/M

State: what your software hands to Jev

The paraphrase says the same thing without the keywords a rules engine looks for.

Questions: the typed decisions you want back

Answers

Run Jev to see typed answers here.

Cascade calculator

A System One model classifies. Code handles the clear majority. A frontier model gets the hard minority. People get the low-confidence residue. Edit the assumptions.

Run log

TimeScenarioInputActionModeLatencyTokensCost

What the independent tests say to keep in mind

  • Threshold on the maximum probability, not the confidence field. Calibrate per question on your own traffic.
  • Yes/no answers ran underconfident, choice and score ran overconfident, on the same inputs. One global threshold is unsafe.
  • The model does not know what it cannot see. Put organisational rules in the state or in code, and give every choice a none_of_these option.
  • Pin a model version before tuning thresholds. Send only the fields the question needs.

Source: The Decision Layer, a CTO briefing on know.2nth.ai.